Anthropic launches Claude Sonnet 4.5, a coding juggernaut

Anthropic has announced Claude Sonnet 4.5, the most powerful coding model the company has released so far, capable of handling complex agentic tasks, gaining reasoning and math improvement over predecessors.

Claude Sonnet 4.5 is a state-of-the-art on SWE-bench Verified evaluation with 82% accuracy, topping Opus 4.1, Sonnet 4, GPT-5 Codex, GPT-5, and Gemini 2.5 Pro. The company has evaluated the model performance, maintaining focus for over 30 hours on complex, multi-step tasks.

Regarding computer use, the Sonnet 4.5 scored 61.4% on real-world computer tasks at the OSWorld benchmark. Besides these, Sonnet 4.5 has visited various benchmark platforms to evaluate its performance against other options, and the result shows a significant boost in product tasks. It has also increased expertise in finance, law, medicine, and STEM compared to Opus 4.1.

Anthropic Claude Sonnet 4.5

Anthropic says that Sonnet 4.5 is released under AI Safety Level 3 (ASL-3) protections, its own framework that matches model capabilities with appropriate safeguards. These include filters called classifiers to detect dangerous inputs and outputs, which can be related to chemical, biological, radiological, and nuclear weapons.

Along with this release, the AI firm is shipping some new upgrades to Claude Code, adding checkpoints to save your progress and enable rollback. It is shipping a new native VS Code extension with a refreshed terminal interface. You can now access the context editing feature and memory tool through the Claude API, allowing agents to run longer and handle more complex tasks.

Claude apps are getting code execution and file creation into the conversation. Furthermore, the Claude for Chrome extension is now available for Max users under the waitlist.

(source)